Gemini Watermark Remover: SynthID Text Defeated

Google began shipping SynthID Text on Gemini output in 2025. By the time GPT-5 launched its own watermarking in early 2026, SynthID had already been live across the Gemini API for nine months — long enough to be the most studied text watermarking system in deployment and the hardest to defeat with off-the-shelf tools.

If you arrived here searching for "Gemini watermark remover", this is the data: five tools tested across a 20-document Gemini corpus, scored against the closest public proxies of Google's own SynthID Text detector, plus three independent text classifiers. Two tools cleared the threshold on most documents. The other three were built for an earlier era of watermarking and have not kept pace.

The same caveat applies as in our ChatGPT watermark remover coverage: we are writing this for commercial Gemini API holders whose use case is platform-distribution clearance. We are not writing it as an academic-integrity evasion guide. If you are a student looking to evade your institution's AI detector, the legal answer is yes and the institutional answer is no, and we are not going to pretend otherwise.

What SynthID Text actually is

The first thing to understand about SynthID Text is that it works differently from the ChatGPT watermark. Both are statistical signatures in the token-distribution of generated text — but the construction is different, and that matters for what defeats them.

ChatGPT's watermark uses probability bias. At each token-generation step, the model identifies tokens that are roughly equally probable, then subtly biases its choice using a hash of the prior context. The bias is small in marginal terms but accumulates across hundreds of tokens. Detectors look for the accumulated bias.

SynthID Text uses tournament sampling. At each token-generation step, the model holds a small tournament among candidate tokens. The winner is chosen using a cryptographic function of the prior context. The output is a watermark embedded in the relationships between consecutive tokens, not just the marginal probability of each token.

The practical implication: SynthID Text survives lighter editing than the ChatGPT scheme. Paraphrasing a paragraph by substituting individual words preserves enough of the inter-token relationships that SynthID's detector can still recognise the watermark. To defeat SynthID, the text needs to be rewritten more aggressively — at the phrase and sentence level, not the word level.

This is why the SynthID Text scores in our benchmark are uniformly lower than the ChatGPT scores for the same tools. The watermark is genuinely harder.

The verdict, before the data

Of the five tools we tested:

1. Undetectable.ai — scored 17 of 20 on our Gemini corpus. Same tool that led our ChatGPT analysis.
2. Humbot — scored 16 of 20. Very close to Undetectable.ai.
3. Manual paraphrase — scored 14 of 20.

The other two (Quillbot's AI humanizer and Stealth Writer) scored 5 and 3 respectively. Both work by substituting individual words rather than rewriting at the structural level; SynthID Text survives those edits.

How we tested

The corpus was 20 documents, ranging from 500 to 1,500 words each, generated across Gemini 2.5 Pro (12 documents) and Gemini 2.0 Flash (8 documents). All documents were generated with realistic prompts: explainers, blog posts, technical writing, marketing copy.

Each document went through each tool. The output was then submitted to four detectors:

• Originality.ai (paid tier, includes SynthID-aware detection as of their late-2025 update)
• GPTZero Pro (general AI detection, included as a cross-model baseline)
• Sapling AI Detector (general AI detection)
• A custom classifier we run that approximates SynthID Text detection using the public research papers Google has released on the watermarking scheme

A document counted as "passed" if all four detectors returned a confidence below 0.3. The 20 in the denominator covers all four detectors per document.

We submitted a small sample (3 documents) to Google's Vertex AI enterprise SynthID detector through an enterprise account we have access to, to validate the public-proxy scoring. Of the 3 documents Undetectable.ai had cleaned, all 3 also passed Google's enterprise detector. Of the 3 raw Gemini documents, all 3 were detected. The sample is small but the directional validation held.

At-a-glance comparison

The 5 tools, ranked

1. Undetectable.ai — the top Gemini watermark remover

Undetectable.ai leads this ranking by margin. Its rewriting pipeline operates at the sentence and clause level — restructuring the text in ways that break SynthID's inter-token relationship patterns, not just substituting vocabulary.

Our score: 17/20 across the 20-document corpus. The three failures clustered around very short documents (under 500 words after rewriting), where the rewriting pass had less material to disrupt the underlying patterns. Documents of 800+ words cleared every detector reliably.

Pricing: $14.99/month for the paid tier (10,000 words/month). Free tier covers ~500 words/day.

Output quality: better than I expected before testing. The rewritten Gemini text reads as competently human-written; complex sentences are simplified appropriately; meaning is preserved across paragraphs.

Verdict: the recommendation for Gemini watermark removal at scale.

2. Humbot — close second

Humbot is one document behind Undetectable.ai (16/20 vs 17/20). The four documents Humbot failed on were also Undetectable.ai failures in two cases, but the differential of two documents (where Undetectable.ai passed and Humbot didn't) is the practical gap.

Humbot's rewriting style is more aggressive — it restructures more, sometimes at the cost of meaning. For technical or terminology-heavy text, this is a liability. For prose where meaning tolerates more flex, it works fine.

Pricing: $14.99/month for the paid tier; ~250 words/day on the free tier.

Verdict: good alternative. If you are mixing ChatGPT and Gemini outputs in a single workflow, either tool works for both.

3. Manual paraphrase — the human-edit fallback

The manual workflow is the same as in the ChatGPT analysis. An experienced editor reads the source text and writes it again in their own words. Score: 14/20.

Why is the score lower for Gemini than for ChatGPT? Because the human editor unconsciously preserves some of the inter-token relationships SynthID embeds — sentence structures, transition patterns, even pacing — even while substituting vocabulary. The editor would need to actively restructure rather than paraphrase to fully scramble SynthID, which requires more time per document.

Verdict: still the best output quality, still expensive in time. For high-stakes documents where the rewriter genuinely understands the content and is willing to restructure, this works. For routine clearance, the automated tools are better.

4. Quillbot AI Humanizer — built for the wrong era

Quillbot was the dominant paraphrasing tool before AI watermarking shipped. The product is well-engineered for its original purpose (smoothing AI-generated text to read more naturally) but it operates at the word-substitution level — exactly the layer that SynthID Text was designed to survive.

Score: 5/20. The five passes were all short documents where word-level substitution happened to disrupt enough of the inter-token relationships by chance.

Verdict: good tool for an earlier era. Not recommended for SynthID Text removal in 2026.

5. Stealth Writer — aggressive but ineffective

Stealth Writer markets itself as a more aggressive humanizer than its peers. The output is indeed more rewritten — significant vocabulary substitution, occasional structural changes. But the rewriting is applied in a way that does not specifically target the patterns SynthID Text screens for.

Score: 3/20. The output also frequently loses meaning or coherence, making the tool unsuitable for serious use beyond the detection-evasion question.

Verdict: not recommended.

Cross-model considerations

A practical question for anyone using multiple LLMs: do the same tools work across ChatGPT, Gemini, and Claude?

Yes, but with different efficacy. Our scores by source model:

The pattern holds: Undetectable.ai > Humbot > Manual > everything else, on every source model. The scores are highest for ChatGPT (which uses the easier-to-defeat probability bias scheme) and lowest for Claude (whose watermarking, recently introduced, is less studied but appears closer to SynthID's construction).

For a mixed-LLM workflow, picking one tool and using it consistently is the right call.

Why Undetectr is not in this list

The same answer as in our ChatGPT watermark remover article: Undetectr is an audio and image tool. It removes the statistical signatures that production music distributors and image platforms screen for. It does not handle text-layer watermarking. The Gemini watermark lives in the token-level structure of the text, which requires text-specific rewriting tooling.

We recommend Undetectr aggressively in the audio category because it is the only tool we have tested that consistently passes the production classifiers. We are not going to bend that recommendation to make the affiliate relationship fit a use case it does not address.

For Gemini watermark removal: Undetectable.ai (first choice) or Humbot. For audio and image artifact removal — the categories most of Artifactr's editorial coverage centres on — Undetectr at $39 one-time for the Lifetime tier remains the recommendation.

The institutional context

This article would be incomplete without the same institutional note we made in our ChatGPT coverage. Most users searching "Gemini watermark remover" in 2026 are not commercial API holders — they are students whose schools use AI detectors. The technical answer to whether SynthID Text can be removed is yes. The institutional answer to whether removing it is permissible is no, regardless of the technical possibility.

The pedagogical landscape will continue to shift. Detectors will improve, humanizers will improve, the cycle will continue. The right institutional response is not better detection but better assignment design — assessments that cannot be replicated by an LLM without the kind of judgement, process documentation, or in-class verification that AI cannot fake. Until that shift happens, every detection-evasion guide on the public internet (including this one) makes the institutional problem harder.

We publish this analysis for the commercial use case it serves and we are honest about what it does for the other use case. That is the best we can offer.

The category outlook for late 2026

Three things we expect to change in the next two quarters:

Google may strengthen SynthID Text. The current implementation can be defeated with $14.99/month tools. Google's published research includes successors to the current scheme — variants that are more resistant to rewriting attacks. When one of those ships, the scores on this page reset.

Detector ensembles will deploy. Currently each watermark scheme has its own detector. The next generation of detectors will test for ChatGPT, Gemini, Claude, and the open-source models simultaneously — making cross-model rewriting strategies less effective at evading any single detector.

Anthropic's watermarking will be documented. Claude has shipped some form of text watermark since late 2025 but the implementation details have not been publicly disclosed at the level OpenAI and Google have published. When Anthropic publishes, we will add Claude-specific coverage to Artifactr.

For now, June 2026: Undetectable.ai. $14.99/month. The honest verdict for Gemini watermark removal in production.